app.py

# Thanks: https://huggingface.co/spaces/stabilityai/stable-diffusion-3-medium
import os
import gradio as gr
import numpy as np
import random
import torch
from diffusers import StableDiffusion3Pipeline, SD3Transformer2DModel, FlowMatchEulerDiscreteScheduler
import spaces
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline

device = "cuda"
dtype = torch.float16

repo = "stabilityai/stable-diffusion-3-medium"
t2i = StableDiffusion3Pipeline.from_pretrained(repo, torch_dtype=torch.float16, revision="refs/pr/26",token=os.environ["TOKEN"]).to(device)

model = AutoModelForCausalLM.from_pretrained(
    "microsoft/Phi-3-mini-4k-instruct", 
    device_map="cuda", 
    torch_dtype=torch.bfloat16, 
    trust_remote_code=True, 
)
tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
upsampler = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
)

generation_args = {
    "max_new_tokens": 300,
    "return_full_text": False,
    "temperature": 0.7,
    "do_sample": True,
    "top_p": 0.95
}

MAX_SEED = np.iinfo(np.int32).max
MAX_IMAGE_SIZE = 1344

@spaces.GPU
def infer(prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps, progress=gr.Progress(track_tqdm=True)):
    messages = [
        {"role": "user", "content": "クールなアニメ風の女の子"},
        {"role": "assistant", "content": "An anime style illustration of a cool-looking teenage girl with an edgy, confident expression. She has piercing eyes, a slight smirk, and colorful hair that flows in the wind. She wears a trendy punk-inspired outfit with a leather jacket, ripped jeans, and combat boots. The background has an urban nighttime feel with city lights and graffiti to match her rebellious vibe. The colors are vibrant with high contrast to give an impactful look. The overall style captures her undeniable coolness and fearless attitude."},
        {"role": "user", "content": prompt },
    ]
    output = upsampler(messages, **generation_args)
    upsampled_prompt=output[0]['generated_text']
    print(upsampled_prompt)
    
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
        
    generator = torch.Generator().manual_seed(seed)
    
    image = t2i(
        prompt = upsampled_prompt, 
        negative_prompt = negative_prompt,
        guidance_scale = guidance_scale, 
        num_inference_steps = num_inference_steps, 
        width = width, 
        height = height,
        generator = generator
    ).images[0] 
    
    return image, seed, upsampled_prompt

examples = [
    "美味しい肉",
    "馬に乗った宇宙飛行士",
    "アニメ風の美少女",
]

css="""
#col-container {
    margin: 0 auto;
    max-width: 580px;
}
"""

with gr.Blocks(css=css) as demo:
    
    with gr.Column(elem_id="col-container"):
        gr.Markdown(f"""
        # 日本語が入力できる SD3 Medium
        """)
        
        with gr.Row():
            
            prompt = gr.Text(
                label="プロンプト",
                show_label=False,
                max_lines=1,
                placeholder="作りたい画像の特徴を入力してください",
                container=False,
            )
            
            run_button = gr.Button("実行", scale=0)
        
        result = gr.Image(label="結果", show_label=False)
        generated_prompt = gr.Textbox(label="生成に使ったプロンプト", show_label=False, interactive=False)
        
        with gr.Accordion("詳細設定", open=False):
            
            negative_prompt = gr.Text(
                label="ネガティブプロンプト",
                max_lines=1,
                placeholder="画像から排除したい要素を入力してください",
            )
            
            seed = gr.Slider(
                label="乱数のシード",
                minimum=0,
                maximum=MAX_SEED,
                step=1,
                value=0,
            )
            
            randomize_seed = gr.Checkbox(label="ランダム生成", value=True)
            
            with gr.Row():
                
                width = gr.Slider(
                    label="横",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=64,
                    value=1024,
                )
                
                height = gr.Slider(
                    label="縦",
                    minimum=256,
                    maximum=MAX_IMAGE_SIZE,
                    step=64,
                    value=1024,
                )
            
            with gr.Row():
                
                guidance_scale = gr.Slider(
                    label="プロンプトの忠実さ",
                    minimum=0.0,
                    maximum=10.0,
                    step=0.1,
                    value=5.0,
                )
                
                num_inference_steps = gr.Slider(
                    label="推論回数",
                    minimum=1,
                    maximum=50,
                    step=1,
                    value=28,
                )
        
        gr.Examples(
            examples = examples,
            inputs = [prompt]
        )
    gr.on(
        triggers=[run_button.click, prompt.submit, negative_prompt.submit],
        fn = infer,
        inputs = [prompt, negative_prompt, seed, randomize_seed, width, height, guidance_scale, num_inference_steps],
        outputs = [result, seed, generated_prompt]
    )

demo.launch()